IRCAM RAVE Model — Neural Audio & Latent Transformations

What this does

IRCAM RAVE Model connects a selected Praat Sound to a TorchScript .ts model. The ordinary processing path calls the model's forward(). For compatible RAVE exports, a separate latent path calls encode(), applies a chosen transformation and optional energy layer, then calls decode().

Key features:

• Four actions — Process sound, Latent transform, Diagnose this model, Diagnose all models in folder.
• Eight latent operations + None — scale, offset, dimension gain, mute, isolate, smooth, freeze, jitter, or leave the encoded latent unchanged.
• Controlled latent energy — deterministic off, reproducible seeded variation, or a new variation on every run.
• Jitter or Smooth drift — absolute latent units or scaled relative to each encoded dimension's spread.
• Model-rate handling — resample to the declared model rate when available, optionally return to the input rate.
• Float audio interchange — Python writes 32-bit IEEE float WAV; normalisation is performed once.
• Model introspection — methods, nn~ *_params buffers, declared sample rate, latent size and compression ratio.

Quick start

1. For Process sound or Latent transform, select exactly one Sound in Praat.
2. Run IRCAM_rave_model.praat.
3. Choose a preset .ts model or Custom, and confirm the models directory.
4. Choose the action. Process sound preserves the established forward-processing path. Latent transform opens a second dialog for latent controls.
5. Leave Normalize = peak for the historical v1.2-style level, or choose none / rms when required.
6. For deterministic latent processing leave Variation mode = Off. Use Seeded or New variation each run to inject controlled latent energy.

Actions

Preset models & detected capabilities

The Praat menu provides seven preset filenames plus a Custom option. The table below summarises the RAVE exports identified by the diagnostics used to develop v1.3–v1.5. Custom models are probed at run time rather than assumed to match these capabilities.

Latent transforms

The latent path operates on the coordinates exactly as returned by the exported model's encode(). Their statistical or semantic interpretation depends on that export.

Latent Energy — controlled instability

The Energy layer is applied after the deterministic latent operation and before decode():

Parameters & advanced settings

Main dialog

Advanced settings

Visualisation

The v1.5.1 figure is designed to show what the algorithm actually did rather than a generic neural-audio illustration.

Diagnostics

Diagnostics require no selected Sound. They load the model and report what the TorchScript export actually exposes, including:

• method names such as forward, encode, decode, prior;
• nn~ buffers such as encode_params, decode_params and forward_params;
• input/output channel counts and rate ratios from four-value method parameter buffers;
• declared sample rate when exposed as an attribute or buffer;
• latent dimension count and compression ratio when encode_params is available.

Technical notes

Normalisation

• none — no scaling or clipping in the Python normaliser. Because the result is float WAV, peaks above full scale can be retained, with a warning.
• peak — scales the result to peak 0.99.
• rms — targets −20 dBFS RMS, attenuating further when required to keep the peak at or below 0.99.
• Gain is applied after the chosen normalisation, so it remains audible.

Sample rate

If the model declares a sample rate and it differs from the Sound, the default path resamples the input to the model rate with a rational windowed-sinc resampler. The result can then be returned to the original input rate or kept at the model rate. If the export does not declare a rate, the script uses the input rate as-is rather than guessing.

Output format

The Python worker writes a 32-bit IEEE float WAV. This avoids the unnecessary PCM16 quantisation used in the older runner and allows Normalize = none to remain genuinely unscaled.

Troubleshooting

• PyTorch not found: install torch and numpy in the Python environment Praat can launch.
• Model file not found: verify the Models directory and selected/custom filename.
• Latent transform rejected: run Diagnose this model; latent processing requires usable encode() and decode().
• CUDA requested but unavailable: choose CPU or Auto.
• Unexpected pitch/time behaviour: inspect the reported sample-rate handling, especially for models with a declared rate different from the input.
• Need the same generative result again: use Seeded variation with the seed reported by the earlier Fresh run.